mutiple fixes: parallel printf, fixed cycle in cache, opencl refactored vecadd and sgemm, regen opencl kernels with hard-float, fixed vortex io bus interface, fixed dpi floats APi to support multicore mode, make vlsim multicore default, make rtlsim multi-core default, removed POCL binaries from repository, updated Makefiles to use external POCL

This commit is contained in:
Blaise Tine
2020-09-19 14:45:42 -04:00
parent 80f929eb61
commit f6f95e0c46
146 changed files with 116779 additions and 194258 deletions

View File

@@ -1,12 +1,16 @@
LLVM_HOME ?= ~/dev/llvm-project/drops
RISCV_TOOLCHAIN_PATH ?= ~/dev/riscv-gnu-toolchain/drops
LLVM_PREFIX ?= /opt/llvm-riscv
RISCV_TOOLCHAIN_PATH ?= /opt/riscv-gnu-toolchain
SYSROOT ?= $(RISCV_TOOLCHAIN_PATH)/riscv32-unknown-elf
POCL_CC_PATH ?= /opt/pocl/compiler
POCL_RT_PATH ?= /opt/pocl/runtime
POCL_CC_PATH ?= $(realpath ../compiler)
POCL_RT_PATH ?= $(realpath ../runtime)
VORTEX_DRV_PATH ?= $(realpath ../../../driver)
VORTEX_RT_PATH ?= $(realpath ../../../runtime)
K_LLCFLAGS += "-O3 -march=riscv32 -target-abi=ilp32f -mcpu=generic-rv32 -mattr=+m,+f -float-abi=hard -code-model=small"
K_CFLAGS += "-v -O3 --sysroot=$(SYSROOT) --gcc-toolchain=$(RISCV_TOOLCHAIN_PATH) -march=rv32imf -mabi=ilp32f -I$(VORTEX_RT_PATH)/include -fno-rtti -fno-exceptions -ffreestanding -nostartfiles -Wl,--gc-sections -Wl,-Bstatic,-T$(VORTEX_RT_PATH)/linker/vx_link.ld"
K_LDFLAGS += "$(VORTEX_RT_PATH)/libvortexrt.a -lm"
CXXFLAGS += -std=c++11 -O0 -g -fpermissive -Wall -Wextra -pedantic -Wfatal-errors
CXXFLAGS += -I$(POCL_RT_PATH)/include
@@ -17,10 +21,10 @@ PROJECT = vecadd
SRCS = main.cc
all: $(PROJECT)
all: $(PROJECT) kernel.pocl
kernel.pocl: kernel.cl
RISCV_TOOLCHAIN_PATH=$(RISCV_TOOLCHAIN_PATH) SYSROOT=$(SYSROOT) LLVM_HOME=$(LLVM_HOME) VORTEX_RT_PATH=$(VORTEX_RT_PATH) POCL_DEBUG=all LD_LIBRARY_PATH=$(LLVM_HOME)/lib:$(POCL_CC_PATH)/lib $(POCL_CC_PATH)/bin/poclcc -o kernel.pocl kernel.cl
LLVM_PREFIX=$(LLVM_PREFIX) POCL_DEBUG=all LD_LIBRARY_PATH=$(LLVM_PREFIX)/lib:$(POCL_CC_PATH)/lib $(POCL_CC_PATH)/bin/poclcc -LLCFLAGS $(K_LLCFLAGS) -CFLAGS $(K_CFLAGS) -LDFLAGS $(K_LDFLAGS) -o kernel.pocl kernel.cl
$(PROJECT): $(SRCS)
$(CXX) $(CXXFLAGS) $^ $(LDFLAGS) -o $@

View File

@@ -1,8 +1,7 @@
kernel void
vecadd (__global const int *a,
__global const int *b,
__global int *c)
__kernel void vecadd (__global const float *A,
__global const float *B,
__global float *C)
{
int gid = get_global_id(0);
c[gid] = a[gid] + b[gid];
C[gid] = A[gid] + B[gid];
}

Binary file not shown.

View File

@@ -1,11 +1,12 @@
#include <stdio.h>
#include <stdlib.h>
#include <assert.h>
#include <math.h>
#include <CL/opencl.h>
#include <string.h>
#include <chrono>
#define SIZE 4096
#define NUM_WORK_GROUPS 2
#define SIZE 256
#define KERNEL_NAME "vecadd"
#define CL_CHECK(_expr) \
@@ -30,20 +31,6 @@
_ret; \
})
int exitcode = 0;
cl_device_id device_id = NULL;
cl_context context = NULL;
cl_command_queue commandQueue = NULL;
cl_program program = NULL;
cl_kernel kernel = NULL;
cl_mem a_memobj = NULL;
cl_mem b_memobj = NULL;
cl_mem c_memobj = NULL;
cl_int *A = NULL;
cl_int *B = NULL;
cl_int *C = NULL;
uint8_t *kernel_bin = NULL;
static int read_kernel_file(const char* filename, uint8_t** data, size_t* size) {
if (nullptr == filename || nullptr == data || 0 == size)
return -1;
@@ -65,6 +52,27 @@ static int read_kernel_file(const char* filename, uint8_t** data, size_t* size)
return 0;
}
static bool almost_equal(float a, float b, int ulp = 4) {
union fi_t { int i; float f; };
fi_t fa, fb;
fa.f = a;
fb.f = b;
return std::abs(fa.i - fb.i) <= ulp;
}
cl_device_id device_id = NULL;
cl_context context = NULL;
cl_command_queue commandQueue = NULL;
cl_program program = NULL;
cl_kernel kernel = NULL;
cl_mem a_memobj = NULL;
cl_mem b_memobj = NULL;
cl_mem c_memobj = NULL;
float *h_a = NULL;
float *h_b = NULL;
float *h_c = NULL;
uint8_t *kernel_bin = NULL;
static void cleanup() {
if (commandQueue) clReleaseCommandQueue(commandQueue);
if (kernel) clReleaseKernel(kernel);
@@ -76,40 +84,15 @@ static void cleanup() {
if (device_id) clReleaseDevice(device_id);
if (kernel_bin) free(kernel_bin);
if (A) free(A);
if (B) free(B);
if (C) free(C);
}
static int find_device(char* name, cl_platform_id platform_id, cl_device_id *device_id) {
cl_device_id device_ids[64];
cl_uint num_devices = 0;
CL_CHECK(clGetDeviceIDs(platform_id, CL_DEVICE_TYPE_ALL, 64, device_ids, &num_devices));
for (int i=0; i<num_devices; i++) {
char buffer[1024];
cl_uint buf_uint;
cl_ulong buf_ulong;
CL_CHECK(clGetDeviceInfo(device_ids[i], CL_DEVICE_NAME, sizeof(buffer), buffer, NULL));
if (0 == strncmp(buffer, name, strlen(name))) {
*device_id = device_ids[i];
return 0;
}
}
return 1;
if (h_a) free(h_a);
if (h_b) free(h_b);
if (h_c) free(h_c);
}
int main (int argc, char **argv) {
printf("enter demo main\n");
cl_platform_id platform_id;
size_t kernel_size;
cl_int binary_status = 0;
int i;
cl_int binary_status;
// read kernel binary from file
if (0 != read_kernel_file("kernel.pocl", &kernel_bin, &kernel_size))
@@ -119,30 +102,19 @@ int main (int argc, char **argv) {
CL_CHECK(clGetPlatformIDs(1, &platform_id, NULL));
CL_CHECK(clGetDeviceIDs(platform_id, CL_DEVICE_TYPE_DEFAULT, 1, &device_id, NULL));
// Creating context.
printf("Create context\n");
context = CL_CHECK2(clCreateContext(NULL, 1, &device_id, NULL, NULL, &_err));
// Memory buffers for each array
a_memobj = CL_CHECK2(clCreateBuffer(context, CL_MEM_READ_ONLY, SIZE * sizeof(cl_int), NULL, &_err));
b_memobj = CL_CHECK2(clCreateBuffer(context, CL_MEM_READ_ONLY, SIZE * sizeof(cl_int), NULL, &_err));
c_memobj = CL_CHECK2(clCreateBuffer(context, CL_MEM_WRITE_ONLY, SIZE * sizeof(cl_int), NULL, &_err));
printf("Allocate device buffers\n");
size_t nbytes = SIZE * sizeof(float);
a_memobj = CL_CHECK2(clCreateBuffer(context, CL_MEM_READ_ONLY, nbytes, NULL, &_err));
b_memobj = CL_CHECK2(clCreateBuffer(context, CL_MEM_READ_ONLY, nbytes, NULL, &_err));
c_memobj = CL_CHECK2(clCreateBuffer(context, CL_MEM_WRITE_ONLY, nbytes, NULL, &_err));
// Allocate memories for input arrays and output arrays.
A = (cl_int*)malloc(sizeof(cl_int)*SIZE);
B = (cl_int*)malloc(sizeof(cl_int)*SIZE);
C = (cl_int*)malloc(sizeof(cl_int)*SIZE);
// Initialize values for array members.
for (i=0; i<SIZE; ++i) {
A[i] = i*2+0;
B[i] = i*2+1;
}
// Create program from kernel source
printf("Create program from kernel source\n");
program = CL_CHECK2(clCreateProgramWithBinary(
context, 1, &device_id, &kernel_size, &kernel_bin, &binary_status, &_err));
if (program == NULL) {
printf("clCreateProgramWithBinary() failed\n");
cleanup();
return -1;
}
@@ -153,42 +125,61 @@ int main (int argc, char **argv) {
// Create kernel
kernel = CL_CHECK2(clCreateKernel(program, KERNEL_NAME, &_err));
// Set arguments for kernel
// Set kernel arguments
CL_CHECK(clSetKernelArg(kernel, 0, sizeof(cl_mem), (void *)&a_memobj));
CL_CHECK(clSetKernelArg(kernel, 1, sizeof(cl_mem), (void *)&b_memobj));
CL_CHECK(clSetKernelArg(kernel, 2, sizeof(cl_mem), (void *)&c_memobj));
// Allocate memories for input arrays and output arrays.
h_a = (float*)malloc(nbytes);
h_b = (float*)malloc(nbytes);
h_c = (float*)malloc(nbytes);
// Initialize values for array members.
for (int i = 0; i < SIZE; ++i) {
h_a[i] = sinf(i)*sinf(i);
h_b[i] = cosf(i)*cosf(i);
h_c[i] = 0xdeadbeef;
//printf("*** [%d]: h_a=%f, h_b=%f\n", i, h_a[i], h_b[i]);
}
// Creating command queue
commandQueue = CL_CHECK2(clCreateCommandQueue(context, device_id, 0, &_err));
commandQueue = CL_CHECK2(clCreateCommandQueue(context, device_id, 0, &_err));
// Copy lists to memory buffers
CL_CHECK(clEnqueueWriteBuffer(commandQueue, a_memobj, CL_TRUE, 0, SIZE * sizeof(float), A, 0, NULL, NULL));
CL_CHECK(clEnqueueWriteBuffer(commandQueue, b_memobj, CL_TRUE, 0, SIZE * sizeof(float), B, 0, NULL, NULL));
printf("Upload source buffers\n");
CL_CHECK(clEnqueueWriteBuffer(commandQueue, a_memobj, CL_TRUE, 0, nbytes, h_a, 0, NULL, NULL));
CL_CHECK(clEnqueueWriteBuffer(commandQueue, b_memobj, CL_TRUE, 0, nbytes, h_b, 0, NULL, NULL));
// Execute the kernel
size_t globalItemSize = SIZE;
size_t localItemSize = SIZE/NUM_WORK_GROUPS;
CL_CHECK(clEnqueueNDRangeKernel(commandQueue, kernel, 1, NULL, &globalItemSize, &localItemSize, 0, NULL, NULL));
printf("Execute the kernel\n");
size_t global_work_size[1] = {SIZE};
size_t local_work_size[1] = {1};
auto time_start = std::chrono::high_resolution_clock::now();
CL_CHECK(clEnqueueNDRangeKernel(commandQueue, kernel, 1, NULL, global_work_size, local_work_size, 0, NULL, NULL));
CL_CHECK(clFinish(commandQueue));
auto time_end = std::chrono::high_resolution_clock::now();
double elapsed = std::chrono::duration_cast<std::chrono::milliseconds>(time_end - time_start).count();
printf("Elapsed time: %lg ms\n", elapsed);
// Read from device back to host.
CL_CHECK(clEnqueueReadBuffer(commandQueue, c_memobj, CL_TRUE, 0, SIZE * sizeof(float), C, 0, NULL, NULL));
printf("Download destination buffer\n");
CL_CHECK(clEnqueueReadBuffer(commandQueue, c_memobj, CL_TRUE, 0, nbytes, h_c, 0, NULL, NULL));
// Test if correct answer
int exitcode = 0;
for (i=0; i<SIZE; ++i) {
if (C[i] != (A[i] + B[i])) {
printf("Failed!\n");
exitcode = 1;
break;
printf("Verify result\n");
int errors = 0;
for (int i = 0; i < SIZE; ++i) {
float ref = h_a[i] + h_b[i];
if (!almost_equal(h_c[i], ref)) {
printf("*** error: [%d] expected=%f, actual=%f, a=%f, b=%f\n", i, ref, h_c[i], h_a[i], h_b[i]);
++errors;
}
}
if (i == SIZE) {
printf("Ok!\n");
if (0 == errors) {
printf("PASSED!\n");
} else {
printf("FAILED! - %d errors\n", errors);
}
// Clean up
cleanup();
return exitcode;
return errors;
}

File diff suppressed because it is too large Load Diff